Feasibility of Interpretable Machine Learning-based

Species Distribution Models: A case of Cormorants in South Korean Rivers

Cheongok Jeon1, Junbeom Bahk2

1The Institute for Korean Regional Studies; LophitaL@snu.ac.kr

2Seoul National University; nongsong@snu.ac.kr

ⓒSaebyeol Yu. Saebyeol’s PowerPoint

Table

•Introduction

•Method

•Result

•Discussion

•Conclusion

•Species distribution model; SDM

> Models for estimating the spatial distribution of species

based on the relationship between the species and environmental variables (Elith and Leathwick, 2009)

Major tool for biogeographic researches as GIS-based environmental information is available

>> Introduction: SDM?

Part 1

https://damariszurell.github.io/SDM-Intro/

Part 1

•Model data is usually built by spatial extraction

•For highly mobile species, if categorical variables are used

without “transformation”, presence information may not mean

“preference”.

•The composition of categorical variables around observed

locations, rather than just at observed locations, can be useful for

capturing the real preference.

Necessity of proper conversion

of categorical variables

Introduction: Issue

A proper transformation is required to reflect the spatial

scale recognized by animals rather than directly utilizing

the data themselves (Guisan and Thuiller, 2005)

Part 1

•SDM studies mainly use numerical data on a macro-scale

(e.g. continental unit).

•Categorical variables can also affect the spatial distribution

of species but are not used properly due to difficulties in

transformation.

•Visually developed taxa, such as aves and primates, can

mainly use the visual landscape for habitat selection.

Estimation of micro-scale spatial distribution

using landscape variables

Introduction: Issue

Part 1

•Machine learning algorithms have

excellent predictive performance due to

their

high flexibility

, but are difficult to

interpret due to their complex function

Prediction accuracy-model interpretability trade-off

•Need to supplement biogeographical

research that requires interpretation of

ecology(Ryo, 2020)

•Attempts to reduce trade-offs and to fit

models that are both predictive and

interpretable by using interpretable

machine learning methods

https://damariszurell.github.io/EEC-MGC/index.html

Introduction: Research Methods

Part 1

Phalacrocorax carbo

•Nearly cosmopolitan diving waterbird

•In 1998, some populations were reported to have

changed as resident birds for the first time(Song, et al. 2017)

•Discussion on the bird control since 2020(Park, 2020)

•Mainly found in isolated islands in large rivers, reservoirs,

and the sea

•Quantitative research to understand the ecology is

insufficient

Phalacrocorax carbo

Phalacrocorax pelagicus

National Institute of Biological Resources

search date:2023-04-21

Introduction: Research Subject

Population Trends of cormorants in South Korea

year

population the number of observation points

Part 2

대충 기술통계 막대그래프

(시간대~출현 자료수)

환경변수 막대그래

•Presence data: 2566 data, 15 sites, 2019-2022

Methods: Species occurrences

The number of presence data

Monitoring year

Geomorphology Vegetation

WATer; WAT

etland Grass Veg

; WGV

River ISland; RIS

etland Wood Veg

; WWV

ArtIficial

Structure

; AIS

Dryland Grass Veg; DGV

FlooDPlain; FDP Dryland Wood Veg; DWV

SanD Dune; SDD

>> Methods: Environment variables

Part 2

•Geomorphic and vegetation

landscape variables

•3km upstream and 5km

downstream for every 15 barrages

•Aerial and drone photography

with field survey

•2-dimensional categorical

variables in spatial polygon

>> Methods: Variable Transformation

Part 2

•Each variable area ratio within the potential landscape perception range was

given as a pixel value

•perception rages:150m, 250m, 400m and 500m

Moving window

Transformation

>> Methods: Modelling

Part 2

•Maximum Entropy; MaxEnt (Philips et al., 2004; 2006)

> No absence data: The study area is limited to near site to the rivers,

which causes the difficulty of determining unsuitable regions

> Estimation of relative habitat suitability using background data representative of a given environment

•We used SHapley Additive exPlanation(SHAP) capable of interpreting each presence data(Lundberg and Lee, 2017)

> Interpretation of our model results using interpretable machine learning method

>> Results: Model Validation

Part 3

•Since there is no absence data for accurate

validation, we validated the models with a

reliability index rather than discriminative index

•The reliability is estimated to be high as the

ratio of presence data giving higher fitness

predictions than random background data is

higher

•Boyce index of 0.9 or higher(high reliability)

(Hirzel et al., 2006)

•In the 250, 500m models, decreases were

confirmed at certain sections, but remained

high overall

Habitat suitability

Perception range

presence/expected ratio

>> Results: Model Prediction

Part 3

•high habitat suitability: where WWV and RIS are adjacent or overlapped

•low habitat suitability: where an area of a specific variable except WAT is too wide

WWV

WAT

RIS

(A) Ipobo

(B) Sejongbo

>> Results: Interpretation(Mean SHAP)

Part 3

•Mean SHAP: the average value of the absolute values of

the Shapley value for the presence data to determine the

variable importance

•Relatively consistent importance with the exception of

the 150m model

•WAT, FDP and RIS show high importance

•WWV contributed relatively consistent

regardless of the models

•Implies that appropriate perception ranges can be applied

per variable

WAT

FDP

RIS

DGV

WWV

SDD

AIS

perception range

Mean SHAP

>> Discussion: Interpretation(SHAP)

Part 4

•a variable that reflects food sources

•If the perception range is reduced, the

area ratio must increase to get the same

contribution = mainly inhabits a body of

water with a certain area

•The water area is also related to the water

depth on the river centerline, implying

food sources and river ice

WAT(water)

image: Flaticon.com

Each point indicates the degree and direction of habitat suitability contribution estimated from

the presence data (Y axis) and the value of variables (X axis)

area ratio

perception range

>> Discussion: Interpretation(SHAP)

Part 4

•a variable that reflects resting spots and

microhabitat

•Area ratios required for positive contributions

are similar among all models = prefer nearby

river islands with proper ‘ratio’

•An overly wide island contributes negatively,

which can refer to narrow bodies of water and

shallow depths

•Even if the island is absent, the negative

contribution is not large, so substitutes may

exist

•However, if present, it can be actively

occupied and make a high contribution

image: Flaticon.com

RIS(river island)

area ratio

perception range

>> Discussion: Interpretation(SHAP)

Part 4

•a variable that represents breeding place

•positive contribution to habitat suitability

regardless of area or ratio

•Locally distributed in the study area, as the

perception range increases, the high-ratio

area decreases, but shows a high positive

contribution

WWV(wetland wood vegetation)

image: Flaticon.com

area ratio

perception range

>> Discussion

Part 4

•A high ratio of WWV can mitigate the negative

contribution of small water areas

•It implies necessity to separate and analyze

breeding sites and food sources, rather than

simply interpreting habitats

Interaction

WAT-WWV

image: Flaticon.com

area ratio of WAT

Shapley value of WAT

area ratio of WWV

>> Discussion

Part 4

•Small artificial infrastructures near water

bodies have a positive contribution

•less avoidance of artificial structures

•Fishways, embankments, and bridge

structures are often used as resting places

(Cho, S. R. and Choi, H. I., 2018)

•There will be no significant impact even in case

of no AIS

•AIS with no WAT is interpreted as a low

contribution

Interaction

AIS-WAT

image: Flaticon.com

area ratio of AIS

Shapley value of AIS

area ratio of WAT

>> Discussion

Part 4

•Occupied as a resting place, similar to AIS, but

with a much more beneficial, high positive

contribution

•SDD near WAT make a positive contribution to

habitat suitability, but its absence doesn’t lead

to negative impacts

Interaction

SDD-WAT

image: Flaticon.com

area ratio of AIS

Shapley value of AIS

area ratio of WAT

>> Discussion

Part 4

•FDP is negatively correlated with WAT

•Too wide FDP is a negative factor due to the difficulty of hunting

•WWV near WAT is preferred but doesn’t show positive

contribution in case of long distance between WWV and WAT

interaction

FDP-WAT, WWV

area ratio of FDP

area ratio of WAT

Shapley value of FDP

area ratio of FDP

area ratio of WWV

Shapley value of FDP

>> Discussion

Part 4

Wetland wood vegetation

River island

Waterbody

•Resting place

•Proper ratio of island and

water is important

•Island boundaries are

mainly used

•Breeding place

•Proper wood area that can

be used as population

breeding

•Mitigating the shortcoming

of narrow water body

•Narrow area is also used

Floodplain

Artificial infrastructure

•Substitute for

river island

•Only those close

to water are used

•The absence is

not critical

•Food source

•Related to water depth and freezing

•Need for certain wide area

•The most important variable

•The main interaction variable with other variables

substitute

•Not significant or negative to

habitat suitability

image: Flaticon.com

>> Conclusion

Part 5

✓Prediction and interpretation of habitat characteristics of cormorants

with geomorphic and vegetation variables

> We suggested that species distribution models can be appropriately fitted at local scales using landscape variables

> We interpreted the landscape use pattern in terms of distance and area ratio

> Using the various perception ranges of a species, we showed that it is important to consider variable landscape use

according to life activities such as spawning and feeding

✓We suggest the possibility of rich ecological interpretation as well as accurate prediction

> Possible to identify complicated ecological interactions, degree and direction of contribution beyond PDP

and permutation-based variable importance

> Using various machine learning algorithms and data, it is possible to explore hypotheses that were not previously

recognized through exploratory model analysis

> Quantitative ecological analysis that supplements qualitative ecological analysis based on expert observations

becomes possible, providing grounds for species conservation policies and increasing academic availability

>> Conclusion

Part 5

✓Simulation-based quantitative verification is required,

and the model needs to be improved using continuous monitoring data

✓Depending on the local context, the range of landscape perception may appear differently

for each barrage

Reference

1. Cho, S. R., & Choi, H. I. (2018) Avifauna and Maintenance Strategy of Gyeongpo Lake, Gangneung.

Korean Journal of Nature Conservation

, 17(1), 45-54.

2. Elith, J., & Leathwick, J. R. (2009) Species distribution models: ecological explanation and prediction across space and time.

Annual review of ecology, evolution, and

systematics

, 40, 677-697.

3. Guisan, A., & Thuiller, W. (2005) Predicting species distribution: offering more than simple habitat models.

Ecology letters

, 8(9), 993-1009.

4. Hirzel, A. H., Le Lay, G., Helfer, V., Randin, C., & Guisan, A. (2006) Evaluating the ability of habitat suitability models to predict species presences.

Ecological modelling

, 199(2)

, 142-152.

5. Lundberg, S. M., & Lee, S. I. (2017) A unified approach to interpreting model predictions.

Advances in neural information processing systems

, 30.

6. Park, J. (2020). 단양강 점령한 가마우지… “물고기 7.5kg 꿀꺽, 그물도 훼손”.

연합뉴스

. https://www.yna.co.kr/view/AKR20200615077100064.

7. Phillips, S. J., Dudík, M., & Schapire, R. E. (2004) A maximum entropy approach to species distribution modeling.

In Proceedings of the twenty-first international conference o

n Machine learning.

655-662.

8. Phillips, S. J., Anderson, R. P., & Schapire, R. E. (2006) Maximum entropy modeling of species geographic distributions.

Ecological modelling

, 190(3-4), 231-259.

9. Ryo, M., Angelov, B., Mammola, S., Kass, J. M., Benito, B. M., & Hartig, F. (2021) Explainable artificial intelligence enhances the ecological interpretability of black‐box species

distribution models.

Ecography

, 44(2), 199-205.

10. Song, H. S., Byeon, J. S., Jo, G. R., Park, J. H., Jeong, J. M., & Park, H. U. (2017) 팔당호에서 민물가마우지의 어류 섭식에 관한 연구.

한국환경과학회 학술발표회 발표논문집,

11, 183-

183.

Thank you!

Jeon, Cheongok1, Bahk, Junbeom2

1The Institute for Korean Regional Studies, Seoul National University, Seoul 08826; LophitaL@snu.ac.kr

2Department of Geography, Seoul National University, Seoul 08826; nongsong@snu.ac.kr

@biogeojeon

cheongokjeon@gmail.com

sleepershark91

nongsong@snu.ac.kr